Fusion of Euclidean Metrics in Featureless Data Analysis: An Equivalent of the Classical Problem of Feature Selection

摘要

The problem of embedding the given set of objects into a linear space with inner product by choosing an appropriate kernel function or set of features is considered to be the major challenge in both kernel-based and feature-based approaches to estimating dependences in data sets of an arbitrary kind. The main idea is to treat several kernels or, more specifically, several numerical features on the same set of objects, as a Cartesian product of the corresponding number of linear spaces, each being supplied with a specific kernel function as a specific inner product.